<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s]]></title><description><![CDATA[<p dir="auto">如题。</p>
<p dir="auto">相较于只有<a href="https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF" rel="nofollow ugc">ISTA-DASLab孤立跑分</a>的IQ3系列，<a href="https://huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4" rel="nofollow ugc">nvfp4有英伟达官方站台</a>，所以其能力会更有说服力。<br />
<img src="https://upload.lcz.me/uploads/29e71445-3bb0-454d-bffc-63f5d6bd6db7.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
<p dir="auto">更何况nvfp4并不是全4bit，很多参数都保持了全量bf16。</p>
<p dir="auto">对于64GB ram + rtx pro 5000 48GB而言。瓶颈主要是因为专家变大了，导致显存只能放进13500个专家，从而专家命中率从99.7%降低至大概97.x~99%。<br />
<img src="https://upload.lcz.me/uploads/8d050e22-3447-4b7d-a806-afba5f272474.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /><br />
<img src="https://upload.lcz.me/uploads/f93f9e6b-ccee-4323-afb6-61a0b35f016a.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
<p dir="auto">然后就是因为内存的紧张，所以Conversation cache只能维系在8GB，4slot。<br />
<img src="https://upload.lcz.me/uploads/6cd49986-cbb6-4267-b6dc-cf769c62432a.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
<p dir="auto">其他的配置保持不变。<br />
如惯例，贴出我的配置表：</p>
<pre><code>args": [
  "--pack",
  "C:\\work\\Strata-data\\models\\Maximilian228--Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata\\pack",
  "--native",
  "C:\\work\\Strata-data\\models\\Maximilian228--Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata\\qwen-nvfp4-gptq-dense.gguf",
  "--native-dense-gguf",
  "C:\\work\\Strata-data\\models\\Maximilian228--Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata\\qwen-nvfp4-gptq-dense.gguf",
  "--ple-gguf",
  "C:\\work\\Strata-data\\models\\Maximilian228--Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata\\ple-fp8.gguf",
  "--embd-gguf",
  "C:\\work\\Strata-data\\models\\Maximilian228--Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata\\token-embd-bf16.gguf",
  "--expert-profile",
  "C:\\work\\Strata\\data\\expert-profile.bin",
  "--expert-cache",
  "auto",
  "--prefill",
  "auto",
  "--spec",
  "4",
  "--mtp",
  "C:\\work\\Strata-data\\mtp\\rt",
  "--max-context",
  "262144",
  "--kv",
  "k8v4",
  "--resident-budget-gib",
  "39",
  "--vision",
  "--vram-reserve-mib",
  "700",
  "--spec-min-p",
  "0.5",
  "--pool-workers",
  "2",
  "--conversation-cache-mib",
  "8192",
  "--conversation-cache-min-tokens",
  "20000"
 ],
</code></pre>
]]></description><link>https://lcz.me/topic/2319</link><generator>RSS for Node</generator><lastBuildDate>Fri, 23 Oct 2026 06:40:36 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2319.rss" rel="self" type="application/rss+xml"/><pubDate>Wed, 14 Oct 2026 05:53:49 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Sat, 17 Oct 2026 14:00:41 GMT]]></title><description><![CDATA[<p dir="auto">向优秀老哥学习选配了rtx pro 5000 48GB + 64GB RAM，目前用llama.cpp 跑qwen3.8 27b Q8 KV 256k 个人用单并发，但是感觉任务一长速度上不去大概decode 30T左右，prefill也是2000多。老哥decode 130t/s，prefill2600 t/s，这个速度非常不错了感觉，不知道flash模型使用质量体感比27b怎么样？</p>
]]></description><link>https://lcz.me/post/25543</link><guid isPermaLink="true">https://lcz.me/post/25543</guid><dc:creator><![CDATA[游志浩]]></dc:creator><pubDate>Sat, 17 Oct 2026 14:00:41 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Thu, 15 Oct 2026 05:44:33 GMT]]></title><description><![CDATA[<p dir="auto">资讯跑得很快。。。。。。。。。我还是需要一点呼吸空间</p>
]]></description><link>https://lcz.me/post/24992</link><guid isPermaLink="true">https://lcz.me/post/24992</guid><dc:creator><![CDATA[imbiplaza ASUS]]></dc:creator><pubDate>Thu, 15 Oct 2026 05:44:33 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 13:08:31 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tony-wang" aria-label="Profile: Tony-Wang">@<bdi>Tony-Wang</bdi></a><br />
我哥你要感谢你的钱包够厚，我推荐人买RTX Pro6000， 也不是所有人都买了。我也知道这卡会涨，但买不起<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f602.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--joy" style="height:1.4em;width:auto;vertical-align:middle" title="😂" alt="😂" /></p>
]]></description><link>https://lcz.me/post/24863</link><guid isPermaLink="true">https://lcz.me/post/24863</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Wed, 14 Oct 2026 13:08:31 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 11:41:39 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a> <a href="/post/24793">说</a>:</p>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E5%85%89%E7%92%9E" aria-label="Profile: 张光璞">@<bdi>张光璞</bdi></a> 其实也不一定不好<br />
2x9700 跟  2x 7900 能做的应该比单张5000 pro 多</p>
<p dir="auto">只是功耗 维护是个问题</p>
</blockquote>
<p dir="auto">N卡的生态还是好太多了。只是多卡玩法比较多样而已.</p>
]]></description><link>https://lcz.me/post/24844</link><guid isPermaLink="true">https://lcz.me/post/24844</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Wed, 14 Oct 2026 11:41:39 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:48:51 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E5%85%89%E7%92%9E" aria-label="Profile: 张光璞">@<bdi>张光璞</bdi></a> 其实也不一定不好<br />
2x9700 跟  2x 7900 能做的应该比单张5000 pro 多</p>
<p dir="auto">只是功耗 维护是个问题</p>
]]></description><link>https://lcz.me/post/24793</link><guid isPermaLink="true">https://lcz.me/post/24793</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:48:51 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:29:28 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tony-wang" aria-label="Profile: Tony-Wang">@<bdi>Tony-Wang</bdi></a> <a href="/post/24786">说</a>:</p>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E5%85%89%E7%92%9E" aria-label="Profile: 张光璞">@<bdi>张光璞</bdi></a>  我是在3.8万时候入手的, 还得感谢 <a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a> 的推荐.</p>
</blockquote>
<p dir="auto">老特配享太庙</p>
]]></description><link>https://lcz.me/post/24787</link><guid isPermaLink="true">https://lcz.me/post/24787</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:29:28 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:26:03 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E5%85%89%E7%92%9E" aria-label="Profile: 张光璞">@<bdi>张光璞</bdi></a>  我是在3.8万时候入手的, 还得感谢 <a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a> 的推荐.</p>
]]></description><link>https://lcz.me/post/24786</link><guid isPermaLink="true">https://lcz.me/post/24786</guid><dc:creator><![CDATA[Tony Wang]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:26:03 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:24:55 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tony-wang" aria-label="Profile: Tony-Wang">@<bdi>Tony-Wang</bdi></a> <a href="/post/24782">说</a>:</p>
<p dir="auto">不过看prefill 2000多好像不太高.</p>
</blockquote>
<p dir="auto">这个还好，我本人用的是claude code驱动，冷启动直接50k token的prefill，总体来讲也能接受。</p>
<p dir="auto">而且这个2600是strata的 bench.bat跑出来的，实际中要高些。</p>
]]></description><link>https://lcz.me/post/24785</link><guid isPermaLink="true">https://lcz.me/post/24785</guid><dc:creator><![CDATA[kop wang]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:24:55 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:24:13 GMT]]></title><description><![CDATA[<p dir="auto">pro5000 真的强。 曾经有一块3万8 的pro5000 摆在我的面前，我没有下单。</p>
<p dir="auto">后来涨到了5万8 ，最后买了2x 9700  2x7900</p>
]]></description><link>https://lcz.me/post/24784</link><guid isPermaLink="true">https://lcz.me/post/24784</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:24:13 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:22:56 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tony-wang" aria-label="Profile: Tony-Wang">@<bdi>Tony-Wang</bdi></a> 那就没有这个烦恼了，Conversation cache可以拉满。</p>
]]></description><link>https://lcz.me/post/24783</link><guid isPermaLink="true">https://lcz.me/post/24783</guid><dc:creator><![CDATA[kop wang]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:22:56 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:22:56 GMT]]></title><description><![CDATA[<p dir="auto">我是 128G <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f609.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--wink" style="height:1.4em;width:auto;vertical-align:middle" title=";)" alt="😉" /> , 不过看prefill 2000多好像不太高.</p>
]]></description><link>https://lcz.me/post/24782</link><guid isPermaLink="true">https://lcz.me/post/24782</guid><dc:creator><![CDATA[Tony Wang]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:22:56 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:21:27 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tony-wang" aria-label="Profile: Tony-Wang">@<bdi>Tony-Wang</bdi></a> 虽然数据上好像比IQ3_XXS掉了很多，但实际使用差距不大。</p>
<p dir="auto">可能跟反复thinking长度变短有关。<br />
最影响体验的就是Conversation cache变少了，导致能缓存的会话变少。</p>
<p dir="auto">如果老哥超过64GB ram，应该可以用到爽</p>
]]></description><link>https://lcz.me/post/24781</link><guid isPermaLink="true">https://lcz.me/post/24781</guid><dc:creator><![CDATA[kop wang]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:21:27 GMT</pubDate></item><item><title><![CDATA[Reply to 你好NVFP4：strata-0.1.44.1，qwen3.8-flash-next，rtx pro 5000 48GB + 64GB RAM，decode 130t/s，prefill2600 t/s on Wed, 14 Oct 2026 06:18:48 GMT]]></title><description><![CDATA[<p dir="auto">先留爪, 准备稍后抄作业 <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f642.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--slightly_smiling_face" style="height:1.4em;width:auto;vertical-align:middle" title=":)" alt="🙂" /></p>
]]></description><link>https://lcz.me/post/24779</link><guid isPermaLink="true">https://lcz.me/post/24779</guid><dc:creator><![CDATA[Tony Wang]]></dc:creator><pubDate>Wed, 14 Oct 2026 06:18:48 GMT</pubDate></item></channel></rss>